TinyFormer: Preserving Tiny Objects in YOLO-DETRHybridReal-time Detectors
Researchers have introduced TinyFormer, a novel hybrid object detection model designed to improve the identification of small objects. This model combines elements of YOLO and DETR architectures, incorporating Vision Transformer representations and a feature pyramid neck. TinyFormer utilizes a Parallel Bi-fusion Module to maintain high-resolution details and a Spatial Semantic Adapter to compensate for spatial information loss in transformer token embeddings. AI
IMPACT Improves accuracy in detecting small objects, potentially benefiting applications like surveillance and autonomous driving.